SLM v1 container overview

Introduces the custom little-endian SLM1 file, section ordering, versioning, supported dtypes, tokenizer choices, tied-output flag, and admission stages.

Experimental
Last verified
2026-06-25 00:00 UTC
Updated
Reading time
2 minutes

Introduces the custom little-endian SLM1 file, section ordering, versioning, supported dtypes, tokenizer choices, tied-output flag, and admission stages.

Implementation evidence: this topic is grounded in the reviewed GGUF.MiRust.com source snapshot. It documents observed code and artifacts without claiming broad deployment, model quality, or production readiness.

Section order

  1. Fixed v1 header.
  2. Tokenizer section: BTOK or BPE1.
  3. Tensor directory containing 64-byte entries.
  4. 64-byte-aligned tensor data and quantization scales.

Format identity

The active implementation is a custom .slm format. It is not GGUF, ONNX, Safetensors, or a standard named SLM format. Compatibility exists only with this parser/writer version.

Supported tensor storage

  • dtype 1: little-endian f32 values.
  • dtype 2: q8_0 signed bytes with one f32 scale per row.
  • dtype 3: q4_0 packed signed nibbles with per-block f32 scales.

Admission stages

Runtime parsing checks structural safety. The native packer validator adds tensor-contract and tokenizer checks. Provenance validation and quality gating are separate layers; passing parser admission does not establish trained quality.

Scope

This starter page defines the questions, boundaries, evidence, and failure modes that should be recorded before a capability is presented as supported.

Engineering considerations

  • Identify the source, version, target environment, and owner.
  • Separate observed values from estimates and externally reported values.
  • Record trade-offs, unsupported cases, and fallback behavior.
  • Link performance statements to a compatible benchmark methodology.

Verification questions

  • What exact artifact, revision, backend, and environment were reviewed?
  • Which assumptions could change the result?
  • Which data should be retained so another engineer can reproduce the conclusion?